Better to Give than to Receive: Predictive Directional Measurement
of Volatility Spillovers
Francis X. Diebold University of Pennsylvania and NBER
Kamil Yilmaz Koç University, Istanbul
First draft/print: November 2008 This draft/print: January 2010
Abstract: Using a generalized vector autoregressive framework in which forecast-error variance decompositions are invariant to variable ordering, we propose measures of both total and directional volatility spillovers. We use our methods to characterize daily volatility spillovers across U.S. stock, bond, foreign exchange and commodities markets, from January 1999 through September 2009. We show that despite significant volatility fluctuations in all four markets during the sample, cross-market volatility spillovers were quite limited until the global financial crisis that began in 2007. As the crisis intensified so too did the volatility spillovers, with particularly important spillovers from the bond market to other markets taking place after the collapse of Lehman Brothers in September 2008. JEL classification numbers: G1, F3 Keywords: Asset Market, Asset Return, Stock Market, Market Linkage, Financial Crisis, Contagion, Vector Autoregression, Variance Decomposition Acknowledgements: For helpful comments we thank two anonymous referees and participants in the Cemapre / IIF International Workshop on the Predictability of Financial Markets, especially Nuno Crato, Antonio Espasa, Antonio Garcia-Ferrer, Raquel Gaspar, and Esther Ruiz. All errors, however, are entirely ours. We thank the National Science Foundation for financial support.
1
1. Introduction
Financial crises occur with notable regularity, and moreover, they display notable
similarities (e.g., Reinhart and Rogoff, 2008). During crises, for example, financial market
volatility generally increases sharply and spills over across markets. One would naturally
like to be able to measure and monitor such spillovers, both to provide “early warning
systems” for emergent crises, and to track the progress of extant crises.
Motivated by such considerations, Diebold and Yilmaz (DY, 2009) introduce a
volatility spillover measure based on forecast error variance decompositions from vector
autoregressions (VARs).1 It can be used to measure spillovers in returns or return volatilities
(or, for that matter, any return characteristic of interest) across individual assets, asset
portfolios, asset markets, etc., both within and across countries, revealing spillover trends,
cycles, bursts, etc. In addition, although it conveys useful information, it nevertheless
sidesteps the contentious issues associated with definition and existence of episodes of
“contagion” or “herd behavior”.2
However, the DY framework as presently developed and implemented has several
limitations, both methodological and substantive. Consider the methodological side. First,
DY relies on Cholesky-factor identification of VARs, so the resulting variance
decompositions can be dependent on variable ordering. One would prefer a spillover
measure invariant to ordering. Second, and crucially, DY addresses only the aggregate
phenomenon of total spillovers (from/to each market i, to/from all other markets, added
1 VAR variance decompositions, introduced by Sims (1980), record how much of the H-step-ahead forecast error variance of some variable, i, is due to innovations in another variable, j. 2 On contagion (or lack thereof) see, for example, Forbes and Rigobon (2002).
2
across i). One would also like to examine directional spillovers (from/to a particular market,
or from a particular market to another market).
Now consider the substantive side. DY considers only the measurement of spillovers
across identical assets (equities) in different countries. But various other possibilities are also
of interest, including individual-asset spillovers within countries (e.g., among the thirty Dow
Jones Industrials in the U.S.), across asset classes (e.g., between stock and bond markets in
the U.S.), and of course various blends. Spillovers across asset classes, in particular, are of
key interest given the global financial crisis that began in 2007 (which appears to have
started in credit markets but spilled over into equities), but they have not yet been
investigated in the DY framework.
In this paper we fill these methodological and substantive gaps. We use a generalized
vector autoregressive framework in which forecast-error variance decompositions are
invariant to variable ordering, and we explicitly include directional volatility spillovers. We
then use our methods in a substantive empirical analysis of daily volatility spillovers across
U.S. stock, bond, foreign exchange and commodities markets over a ten year period,
including the recent global financial crisis.
We proceed as follows. In section 2 we discuss our methodological approach,
emphasizing in particular our new use of generalized variance decompositions and
directional spillovers. In section 3 we describe our data and present our substantive results.
We conclude in section 4.
3
2. Methods: Generalized Spillover Definition and Measurement
Here we extend the DY spillover index, which follows directly from the familiar
notion of a variance decomposition associated with an N-variable vector autoregression.
Whereas DY focuses on total spillovers in a simple VAR framework (i.e., with potentially
order-dependent results driven by Cholesky factor orthogonalization), we progress by
measuring directional spillovers in a generalized VAR framework that eliminates the
possible dependence of results on ordering.
Consider a covariance stationary N-variable VAR(p), 1
p
t i t i ti
x x ε−=
= Φ +∑ , where
(0, )ε Σ∼ is the vector of independently and identically distributed disturbances. The
moving average representation is 0
t i t ii
x Aε∞
−=
= ∑ , where the NxN coefficient matrices iA obey
the recursion 1 1 2 2 ...i i i p i pA A A A− − −= Φ + Φ + + Φ , with 0A an NxN identity matrix and 0iA =
for i<0. The moving average coefficients (or transformations such as impulse-response
functions or variance decompositions) are the key to understanding the dynamics of the
system. We rely on variance decompositions, which allow us to parse the forecast error
variances of each variable into parts attributable to the various system shocks. Variance
decompositions allow us to assess the fraction of the H-step-ahead error variance in
forecasting ix that is due to shocks to ,jx j i∀ ≠ , for each i.
Calculation of variance decompositions requires orthogonal innovations, whereas our
VAR innovations are generally contemporaneously correlated. Identification schemes such
as that based on Cholesky factorization achieve orthogonality, but the variance
decompositions then depend on ordering of the variables. We circumvent this problem by
4
exploiting the generalized VAR framework of Koop, Pesaran and Potter (1996) and Pesaran
and Shin (1998), hereafter KPPS, which produces variance decompositions invariant to
ordering. Instead of shocking all variables of the system at once and orthogonalizing these
shocks through Cholesky decomposition or through the structural VAR approach, the
generalized VAR approach shocks only one variable at a time and integrates out the effects
of other shocks using the historically observed distribution of the errors. As the shocks to
each variable are not orthogonalized, the sum of contributions to the variance of forecast
error (that is, the row sum of the elements of the variance decomposition table) is not
necessarily equal to one.
Variance Shares
Let us define own variance shares to be the fractions of the H-step-ahead error
variances in forecasting ix due to shocks to ix , for i=1, 2,..,N, and cross variance shares, or
spillovers, to be the fractions of the H-step-ahead error variances in forecasting ix due to
shocks to jx , for i, j = 1, 2,.., N, such that i j≠ .
Denoting the KPPS H-step-ahead forecast error variance decompositions by ( )gij Hθ ,
for H = 1, 2, ..., we have
11 ' 20
1 ' '0
( )( )
( )
Hii i h jg h
ij Hi h h ih
e A eH
e A A e
σθ
−−=
−
=
= ∑ ∑∑ ∑
(1)
where ∑ is the variance matrix for the error vector ε , iiσ is the standard deviation of the
error term for the ith equation and ie is the selection vector with one as the ith element and
zeros otherwise. As explained above, the sum of the elements of each row of the variance
5
decomposition table is not equal to 1: 1
( ) 1N
gij
j
Hθ=
≠∑ . In order to use the information
available in the variance decomposition matrix in the calculation of the spillover index, we
normalize each entry of the variance decomposition matrix by the row sum as3:
1
( )( )
( )
gijg
ij Ng
ijj
HH
H
θθ
θ=
=
∑ (2)
Note that, by construction, 1
( ) 1N
gij
jHθ
=
=∑ and , 1
( )N
gij
i jH Nθ
=
=∑ .
Total Spillovers
Using the volatility contributions from the KPPS variance decomposition, we can
construct a total volatility spillover index:
, 1 , 1
, 1
( ) ( )
( ) 100 100( )
N Ng g
ij iji j i ji j i jg
Ng
iji j
H H
S HNH
θ θ
θ
= =≠ ≠
=
= =
∑ ∑
∑i i . (3)
This is the KPPS analog of the Cholesky factor based measure used in Diebold and Yilmaz
(2009). The total spillover index measures the contribution of spillovers of volatility shocks
across four asset classes to the total forecast error variance.
Directional Spillovers
Although it is sufficient to study the total volatility spillover index to understand how
much of shocks to volatility spill over across major asset classes, the generalized VAR 3 Alternatively, we can normalize the elements of the variance decomposition matrix with the column sum of these elements and compare the resulting total spillover index with the one obtained from the normalization with the row sum.
6
approach enables us to learn about the direction of volatility spillovers across major asset
classes. As the generalized impulse responses and variance decompositions are invariant to
the ordering of variables, we calculate the directional spillovers using the normalized
elements of the generalized variance decomposition matrix. We measure directional volatility
spillovers received by market i from all other markets j as:
1
1
( )
( ) 100( )
Ng
ijjj ig
i Ng
ijj
H
S HH
θ
θ
=≠
=
=
∑
∑i i (4)
In similar fashion we measure directional volatility spillovers transmitted by market i to all
other markets j as:
1
1
( )
( ) 100( )
Ngji
jj ig
i Ngji
j
H
S HH
θ
θ
=≠
=
=
∑
∑i i (5)
One can think of the set of directional spillovers as providing a decomposition of total
spillovers into those coming from (or to) a particular source.
Net Spillovers
Finally, we obtain the net volatility spillover from market i to all other markets j as
( ) ( ) ( )g g gi i iS H S H S H= −i i (6)
The net volatility spillover is simply the difference between gross volatility shocks
transmitted to and gross volatility shocks received from all other markets.
Net Pairwise Spillovers
7
The net volatility spillover (6) provides summary information about how much in net
terms each market contributes to volatility in other markets. It is also of interest to examine
net pairwise volatility spillovers, which we define as:
1 1
( ) ( )( ) 100
( ) ( )
g gij jig
ij N Ng g
ik jkk k
H HS H
H H
θ θ
θ θ= =
⎛ ⎞⎜ ⎟⎜ ⎟= −⎜ ⎟⎜ ⎟⎝ ⎠∑ ∑
i (7)
The net pairwise volatility spillover between markets i and j is simply the difference between
gross volatility shocks transmitted from market i to j and gross volatility shocks transmitted
from j to i.
3. Empirics: Estimates of Volatility Spillovers across U.S. Asset Markets
Here we use our framework to measure volatility spillovers among four key U.S.
asset classes: stocks, bonds, foreign exchange and commodities. This is of particular interest
because spillovers across asset classes may be an important aspect of the global financial
crisis that began in 2007 (which started in credit markets but spilled over into equities).
In the remainder of this section we proceed as follows. We begin by describing our
data in section 3a. Then we calculate average (i.e., total) spillovers in section 3b. We then
quantify spillover dynamics, examining rolling-sample total spillovers, rolling-sample
directional spillovers, rolling-sample net directional spillovers and rolling-sample net
pairwise spillovers below.
Stock, Bond, Exchange Rate, and Commodity Market Volatility Data
We examine daily volatilities of returns on U.S. stock, bond, foreign exchange, and
commodity markets. In particular, we examine the S&P 500 index, the 10-year Treasury
8
bond yield, the New York Board of Trade U.S. dollar index futures, and the Dow-Jones/UBS
commodity index.4 The data span January 25, 1999 through September 30, 2009, for a total
of 2688 daily observations.
In the tradition of a large literature dating at least to Parkinson (1980), we estimate
daily variance using daily high and low prices.5 For market i on day t we have
22 max min0.361 ln( ) ln( )it it itP Pσ ⎡ ⎤= −⎣ ⎦ ,
where maxitP is the maximum (high) price in market i on day t, and min
itP is the daily minimum
(low) price. Because 2itσ is an estimator of the daily variance, the corresponding estimate of
the annualized daily percent standard deviation (volatility) is 2ˆ 100 365it itσ σ= • . We plot
the four markets’ volatilities in Figure 1 and we provide summary statistics in Table 1.
Several interesting facts emerge, including: (1) The bond and stock markets have been the
most volatile (roughly equally so), with commodity and FX markets comparatively less
volatile, (2) volatility dynamics appear highly persistent, in keeping with a large literature
summarized for example in Andersen, Bollerslev, Christoffersen and Diebold (2006), and (3)
all volatilities are high during the recent crisis, with stock and bond market volatility, in
particular, displaying huge jumps.
Throughout the sample, stock market went through two periods of major volatility.
In 1999, daily stock market volatility was close to 25 percent, but it increased significantly to
fluctuate above 25 percent until mid-2003, moving occasionally above 50 percent. After
4 The DJ/AIG commodity index was re-branded as the DJ/UBS commodity index following the acquisition of AIG Financial Products Corp. by UBS Securities LLC on May 6, 2009. 5 For background, see Alizadeh, Brandt and Diebold (2002) and the references therein.
9
mid-2003, it declined to less than 25 percent and stayed there until August 2007. Since
August 2007, stock market volatility reflects well the dynamics of the sub-prime crisis.
In the first half of our sample, the interest rate volatility measured by the annualized
standard deviation was comparable to the stock market volatility. While it was lower than 25
percent mark for most of 2000, in the first and last few months of 2001, it increased and
fluctuated between 25-50 percent. Bond market volatility remained high until mid-2005, and
fell below 25 percent from late 2005 through the first half of 2007. Since August 2007,
volatility in bond markets has also increased significantly.
Commodity market volatility used to be very low compared to stock and bond
markets, but it increased slightly over time and especially in 2005-2006 and recently in 2008.
FX market volatility has been the lowest among the four markets. It increased in 2008 and
moved to a 25-50 percent band following the collapse of Lehman Brothers in September
2008. Since then, FX market volatility declined, but it is still above its average for the last
decade.
Unconditional Patterns: The Full-Sample Volatility Spillover Table
We call Table 2 a volatility spillover table. Its thij entry is the estimated contribution
to the forecast error variance of market i coming from innovations to market j.6 Hence the
off-diagonal column sums (labeled contributions to others) or row sums (labeled
contributions from others), are the “to” and “from” directional spillovers, and the “from
minus to” differences are the net volatility spillovers. In addition, the total volatility spillover
6 All results are based on vector autoregressions of order 4 and generalized variance decompositions of 10-day-ahead volatility forecast errors. To check for the sensitivity of the results to the choice of the order of VAR we calculate the spillover index for orders 2 through 6, and plot the minimum, the maximum and the median values obtained in Figure A1 of the Appendix. Similarly, we calculated the spillover index for forecast horizons varying from 4 days to 10 days. Both Figure A1 and Figure A2 of the Appendix show that the total spillover plot is not sensitive to the choice of the order of VAR or to the choice of the forecast horizon.
10
index appears in the lower right corner of the spillover table. It is approximately the grand
off-diagonal column sum (or row sum) relative to the grand column sum including diagonals
(or row sum including diagonals), expressed as a percent.7 The volatility spillover table
provides an approximate “input-output” decomposition of the total volatility spillover index.
Consider first what we learn from the table about directional spillovers (gross and
net). From the “directional to others” row, we see that gross directional volatility spillovers to
others from each of the four markets are not very different. We also see from the “directional
from others” column that gross directional volatility spillovers from others to FX is relatively
large, at 26.53 percent, followed by the bond market with the spillovers from others
explaining 17.21 percent of the forecast error variance. As for net directional volatility
spillovers, the largest are from the stock market to others (6.92 percent) and from others to
the FX market (10.27 percent).
Now consider the total (non-directional) volatility spillover, which is effectively a
distillation of the various directional volatility spillovers into a single index. The total
volatility spillover appears in the lower right corner of Table 2, which indicates that on
average, across our entire sample, 16.55 percent of volatility forecast error variance in all
four markets comes from spillovers.
Conditioning and Dynamics I: The Rolling-Sample Total Volatility Spillover Plot
Clearly, many changes took place during the years in our sample, January 1999-
September 2009. Some are well-described as more-or-less continuous evolution, such as
increased linkages among global financial markets and increased mobility of capital, due to 7 As we have already discussed in Section 2 in detail, the approximate nature of the claim stems from the properties of the generalized variance decomposition. With Cholesky factor identification the claim is exact rather than approximate; see also Diebold and Yilmaz (2009).
11
globalization, the move to electronic trading, and the rise of hedge funds. Others, however,
may be better described as bursts that subsequently subside.
Given this background of financial market evolution and turbulence, it seems unlikely
that any single fixed-parameter model would apply over the entire sample. Hence the full-
sample spillover table and spillover index constructed earlier, although providing a useful
summary of “average” volatility spillover behavior, likely miss potentially important secular
and cyclical movements in spillovers. To address this issue, we now estimate volatility
spillovers using 100-day rolling samples, and we assess the extent and the nature of spillover
variation over time via the corresponding time series of spillover indices, which we examine
graphically in the so-called total spillover plot of Figure 2.
Starting around twenty percent in 1999, the total volatility spillover plot usually
fluctuates between ten and thirty percent. However, there are important exceptions: The
spillovers exceed the thirty percent mark in the second half of 2000 and the first quarter of
2001, immediately after the 9/11 terrorist attacks, in the third quarter of 2002, in June 2006
and most importantly by far, during the global financial crisis of 2007-2009. One can see
five volatility waves during the recent crisis: July-August 2007 (credit crunch), January 2008
(panic in stock and foreign exchange markets followed by an unscheduled rate cut of three-
quarters of a percentage points by Federal Reserve), June 2008 (when the worries about the
burden of the crisis on government budgets started to appear, the volatility increased in the
bond market and then spread to the other markets), following the collapse of Lehman
Brothers (September-October 2008) and in the first half of 2009 as the crisis started to have
its real effects on the world economy. During these episodes, the spillover index surges well
above thirty percent. Indeed, following the collapse of Lehman Brothers in mid-September,
12
and consistent with the unprecedented evaporation of liquidity world-wide, the volatility
spillover plot jumped to 64 percent on September 30, 2008, before declining slightly.
Conditioning and Dynamics II: Rolling-Sample Gross Directional Volatility Spillover Plots
Thus far we have discussed the total spillover plot, which is of interest but discards
directional information. That information is contained in the “Contribution to” row (the sum
of which is given by ( )giS Hi in equation 4) and the “Contribution from” column (the sum of
which is given by ( )giS Hi in equation 5).
We now estimate that row and column dynamically, in a fashion precisely parallel to
the earlier-discussed total spillover plot. We call these directional spillover plots. In Figure
3, we present the directional volatility spillovers from our four asset classes. They vary
greatly over time. During tranquil times, spillovers from each market are below five percent,
but during volatile times, directional spillovers increase to above 10 percent. Among the four
markets, gross volatility spillovers from the commodity markets are in generally smaller than
the spillovers from the other three markets.
In Figure 4, we present directional volatility spillovers to our four asset classes. As
with the directional spillovers from others, the spillovers to others vary noticeably over time.
The relative variation pattern, however, is reversed, with directional volatility spillovers to
commodities and FX increasing relatively more in turbulent times.
Conditioning and Dynamics III: Rolling-Sample Net Directional Volatility Spillover Plots
Above we briefly discussed the gross spillover plots, because our main focus point is
the net directional spillover plot, which is given in Figure 5. Each point in Figure 5a through
5d corresponds ( )giS H (equation 6) and is the difference between the “Contribution from”
13
column sum and the “Contribution to” row sum. In addition, as we described briefly at the
end of section 2, we also calculate net pairwise spillovers between two markets (equation 7)
and present these plots in Figure 6.
First note that, overall, there has been relatively little net volatility transmission from
the commodity and FX markets. Only in the first five months of 2003 (before and
immediately after the invasion of Iraq in March 2003), in late 2004 and early 2005 and
during the commodity price boom in the first half of 2008 do we observe that net volatility
spillovers from commodity markets to other markets reach five percent. Similarly, volatility
in FX markets also had very little net impact on volatility in other markets, perhaps with the
exception of the sizable spillovers at the end of 2001, throughout 2006 and in January 2008.
Instead, the clear channels of net directional volatility spillovers are from the stock
and bond markets. Net volatility spillovers from the stock market for the most part appear to
be positive and large especially in the earlier parts of the sample period (from 2000 to 2002,
in particular). As the technology bubble burst in March 2000, net volatility spillovers from
the stock market reached close to 5 percent (Figure 5a). At the time, the bulk of the volatility
spillovers from the stock market went to the bond market (Figure 6a). However, the stock
market volatility and the spillovers from the stock market continued to run high towards the
end of 2000 and in 2001, as the problems in technology stocks spread to other stocks.
Volatility spillovers from the stock market ran as high as 3-4 percent from the end of
2000 until the 9/11 terrorist attacks in September (Figure 5a). During this period, shocks to
stock market volatility were transmitted to the commodities and the FX markets. After the
terrorist attacks on September 11, 2001, volatility spillovers from the stock market took place
only in the direction of the commodity market. During the increased U.S. stock market
14
gyrations in June through October 2002, net spillovers from the stock market reached close
to 10 percent, but affected only the FX and bond markets. Finally, from August 2007 to June
2008, net spillovers from the stock market to bond and commodity markets reached close to
5 percent mark. While volatility spillovers from the stock market affected the FX market
after the collapse of Lehman Brothers, during the panic in January 2008 it was the stock
market that received volatility spillovers from the FX market.
Similarly, and interestingly, during the global financial crisis and especially since the
summer of 2008, the bond market was the most important transmitter of volatility. Indeed, in
June 2008 when the worries about the heavy burden of the crisis on government finances
intensified, and following the collapse of Lehman Brothers in mid-September, net volatility
spillovers from the bond market reached as high as 10 percent (Figure 5b). The bulk of the
volatility spillovers from the bond market during this time period were transmitted to the
stock market (see Figure 6a). Over the same period, the commodity and FX markets were
also net receivers of volatility spillovers from the bond market (Figure 6d and 6e).
Volatility spillovers from commodity markets increased in 2003 just before the
invasion of Iraq by U.S. forces, at the end of 2004 and early 2005, when the surge in Chinese
demand for oil and metals surprised investors sending commodity prices higher (shocks
mostly transmitted to the bond and FX markets), and especially during the commodity price
boom in the first half of 2008 (shocks mostly transmitted to the bond market). The volatility
shocks in the commodity market before and during the initial phases of the Iraqi invasion
spilled over mostly to the stock market, but also to the bond market. During the late 2004-
early 2005 and the first half of 2008, the volatility shocks in the commodity market mostly
spilled over to the bond market, but also to the FX market.
15
Finally, volatility spillovers from the FX market increased at the end of 2001 and
early 2002. It also increased in May 2006, following the FED’s decision to increase the
Federal Funds target rate further to 5 percent on May 10, 2006, and in January 2008,
following the FED’s decision to cut the Federal Funds target rate by 75 basis points on
January 22, 2008. In all three episodes, the volatility shocks in the FX market spilled over to
the stock market. In only two out of these three episodes (May 2006 and January 2008)
shocks to FX return volatility spilled over to the bond market, albeit at a lower scale.
Our results show that when there is a significant shock to volatility in one of the four
major asset markets, this volatility shock is likely to spill over to other asset markets. Before
the global financial crisis of 2008-2009 mostly this was the case; shocks did not occur
simultaneously. When one of the markets was hit by a volatility shock, the shock was likely
to spill over to other markets if it was significant enough. As a case in point we can have a
closer look at volatility spillovers in 2006. When the Federal Reserve decided to increase the
FED Funds target rate from 4.75 percent to 5.00 percent in May 2006 and signaled it might
increase it further in June, it had a rather unexpected effect on FX markets. The dollar
started to gain momentum as the investors decided to move part of the portfolio they invested
in many emerging markets back to the US safe haven. During this period the volatility in FX
markets spilled over to other markets, but most importantly to the commodity market.
We know that during the global financial crisis volatility shocks take place
simultaneously in some or all asset markets. In such a case it is not possible to argue that the
volatility spillovers will always take place from a particular market towards the others. For
this we can focus on January 2008 episode. In an unscheduled emergency meeting on
January 21, 2008, which was an official holiday in the US, the Federal Reserve’s FOMC
16
decided to lower Federal Funds target rate by three-quarters of a percent. Before that day
volatility was high in all asset markets. After this meeting volatility increased further because
it was taken as an acknowledgement by the Federal Reserve of how grave the situation was.
While the volatility in the stock and bond markets increased by 30 and 123 percent,
respectively, on the next business day (January 22), volatility in the commodity and FX
markets increased by 186 and 174 percent. Furthermore, in the following weeks investors
moved away from other assets towards commodities as they were deemed less vulnerable
during the financial crisis. In addition investors also reduced their US dollar holdings while
increasing that of Euro and Yen. As a result following the January 21, 2008 decision the
volatility shocks were transmitted from the FX market to the stock and bond markets,
whereas shocks in the commodity markets were transmitted to the bond market.
4. Concluding Remarks
We have provided both gross and net directional spillover measures that are independent of
the ordering used for volatility forecast error variance decompositions. When applied to U.S.
financial markets, our measures shed new light on the nature of cross-market volatility
transmission, pinpointing the importance during the recent crisis of volatility spillovers from
the bond market to other markets.
We are of course not the first to consider issues related to volatility spillovers (e.g.,
Engle et al. 1990; King et al., 1994; Edwards and Susmel, 2001), but our approach is very
different. It produces continuously-varying indexes (unlike, for example, the “high state /
low state” indicator of Edwards and Susmel), and it is econometrically tractable even for
very large numbers of assets. Although it is beyond the scope of this paper, it will be
interesting in future work to understand better the relationship of our spillover measure to a
17
variety of others based on measures ranging from traditional (albeit time-varying)
correlations (e.g., Engle, 2002, 2009) to the recently-introduced CoVaR of Adrian and
Brunnermeier (2008).
18
References
Adrian, T. and Brunnermeier, M. (2008), “CoVaR,” Staff Report 348, Federal Reserve
Bank of New York.
Alizadeh, S., Brandt, M.W. and Diebold, F.X. (2002), “Range-Based Estimation
of Stochastic Volatility Models,” Journal of Finance, 57, 1047-1092.
Andersen, T.G., Bollerslev, T., Christoffersen, P.F. and Diebold, F.X. (2006), “Practical
Volatility and Correlation Modeling for Financial Market Risk Management,” in M.
Carey and R. Stulz (eds.), Risks of Financial Institutions, University of Chicago Press
for NBER, 513-548.
Diebold, F.X. and Yilmaz, K. (2009), “Measuring Financial Asset Return and Volatility
Spillovers, With Application to Global Equity Markets,” Economic Journal, 119,
158-171.
Engle, R.F. (2002), “Dynamic Conditional Correlation: A Simple Class of Multivariate
GARCH Models,” Journal of Business and Economic Statistics, 20, 339-350.
Engle, R.F. (2009), Anticipating Correlations. Princeton: Princeton University Press.
Engle, R.F., Ito, T. and Lin, W.-L. (1990), “Meteor Showers or Heat Waves:
Heteroskedastic Intra-Daily Volatility in the Foreign Exchange Market,”
Econometrica, 58, 525-542.
Forbes, K.J. and Rigobon, R. (2002), “No Contagion, Only Interdependence:
Measuring Stock Market Comovements,” Journal of Finance, 57, 2223-2261.
King, M., Sentana, E. and Wadhwani, S. (1994), “Volatility and Links Between
National Stock Markets,” Econometrica, 62, 901-933.
Koop, G., Pesaran, M.H., and Potter, S.M. (1996), “Impulse Response Analysis in Non-
Linear Multivariate Models,” Journal of Econometrics, 74, 119–147.
19
Parkinson, M. (1980), “The Extreme Value Method for Estimating the Variance of the Rate
of Return,” Journal of Business, 53, 61-65.
Pesaran, M.H. and Shin, Y. (1998), “Generalized Impulse Response Analysis in Linear
Multivariate Models,” Economics Letters, 58, 17-29.
Reinhart, C.M. and Rogoff, K.S. (2008), “Is the 2007 U.S. Subprime Crisis So Different? An
International Historical Comparison,” American Economic Review, 98, 339–344.
Sims, C.A. (1980), “Macroeconomics and Reality,” Econometrica, 48, 1-48.
20
Figure 1. Daily U.S. Financial Market Volatilities
(Annualized Standard Deviation, Percent)
a) Stock Market - S&P500
0
25
50
75
100
125
150
99 00 01 02 03 04 05 06 07 08 090
25
50
75
100
125
150
99 00 01 02 03 04 05 06 07 08 09
b) Bond Market - 10-year Interest Rate
0
25
50
75
100
125
150
99 00 01 02 03 04 05 06 07 08 09
c) Commodity Market - DJ/AIG Index
0
25
50
75
100
125
150
99 00 01 02 03 04 05 06 07 08 09
d) FX Market - US Dollar Index Futures
21
Table 1: Volatility Summary Statistics, Four Asset Classes
Stocks Bonds Commodities FX Mean 18.05 20.34 11.57 8.74
Median 14.71 16.88 9.78 7.79
Maximum 125.17 230.77 80.11 42.34
Minimum 2.75 1.94 0.20 0.42
Std. Deviation 12.93 13.88 8.21 4.52
Skewness 2.98 3.11 1.66 1.82
Kurtosis 17.71 28.60 7.66 9.23
Table 2: Volatility Spillover Table, Four Asset Classes
Stocks Bonds Commodities FX Directional FROM Others
Stocks 90.10 4.92 3.26 1.73 9.90
Bonds 2.95 82.79 4.20 10.06 17.21
Commodities 5.43 2.64 87.45 4.48 12.55
FX 8.44 11.34 6.75 73.47 26.53
Directional TO Others 16.82 18.90 14.21 16.26 Directional Including Own 106.9 101.7 101.7 89.7 Total Spillover
Index (66.19/400): 16.6%
22
Figure 2. Total Volatility Spillovers, Four Asset Classes
0
10
20
30
40
50
60
70
99 00 01 02 03 04 05 06 07 08 09
23
Figure 3. Directional Volatility Spillovers, FROM Four Asset Classes
0
5
10
15
20
25
30
99 00 01 02 03 04 05 06 07 08 09
a) Stock Market - S&P500
0
5
10
15
20
25
30
99 00 01 02 03 04 05 06 07 08 09
b) Bond Market - 10-year Interest Rate
0
5
10
15
20
25
30
99 00 01 02 03 04 05 06 07 08 09
c) Commodity Market - DJ/UBS Commodity Index
0
5
10
15
20
25
30
99 00 01 02 03 04 05 06 07 08 09
d) FX Market - US Dollar Index Futures
24
Figure 4. Directional Volatility Spillovers, TO Four Asset Classes
0
4
8
12
16
20
99 00 01 02 03 04 05 06 07 08 09
a) Stock Market - S&P500
0
4
8
12
16
20
99 00 01 02 03 04 05 06 07 08 09
b) Bond Market - 10-year Interest Rate
0
4
8
12
16
20
99 00 01 02 03 04 05 06 07 08 09
c) Commodity Market - DJ/UBS Commodity Index
0
4
8
12
16
20
99 00 01 02 03 04 05 06 07 08 09
d) FX Market - US Dollar Index Futures
25
Figure 5. Net Volatility Spillovers, Four Asset Classes
-10
0
10
20
99 00 01 02 03 04 05 06 07 08 09
a) Stock Market - S&P500
-10
0
10
20
99 00 01 02 03 04 05 06 07 08 09
b) Bond Market - 10-year Interest rate
-10
0
10
20
99 00 01 02 03 04 05 06 07 08 09
c) Commodity market - DJ/UBS Commodity Index
-10
0
10
20
99 00 01 02 03 04 05 06 07 08 09
d) FX Market - US Dollar Index Futures
26
Figure 6. Net Pairwise Volatility Spillovers
-10
-5
0
5
10
15
99 00 01 02 03 04 05 06 07 08 09
a) Stock Market - Bond Market
-10
-5
0
5
10
15
99 00 01 02 03 04 05 06 07 08 09
b) Stock Market - Commodity Market
-10
-5
0
5
10
15
99 00 01 02 03 04 05 06 07 08 09
c) Stock Market - FX Market
-10
-5
0
5
10
15
99 00 01 02 03 04 05 06 07 08 09
d) Bond Market - Commodity Market
-10
-5
0
5
10
15
99 00 01 02 03 04 05 06 07 08 09
e) Bond Market - FX Market
-10
-5
0
5
10
15
99 00 01 02 03 04 05 06 07 08 09
f) Commodity Market - FX Market
27
APPENDIX Figure A1. Sensitivity of the index to VAR lag structure (Max, Min and Median values
of the index for VAR order of 2 through 6)
0
10
20
30
40
50
60
99 00 01 02 03 04 05 06 07 08 09
Median (Max,Min) Figure A2. Sensitivity of the Index to Forecast Horizon (Min, Max and Median values
over 5 to 10-day horizons)
0
10
20
30
40
50
60
99 00 01 02 03 04 05 06 07 08 09Median (Max,Min)